The Day in AI

• DeepSeek added Search and Fetch endpoints to its API on 17 October, each at 5 yuan per 1,000 successful requests and 5 queries per second.
• Kyutai released REKI, a 2.4 billion parameter robot model that remembers its last three camera frames for an extra 0.1 ms per step.
• Meta released Muse Spark Open on 15 October, an 816 billion parameter open-weight model that it calls the most capable US open-weight model.
• Disney agreed to hire more than 60 Character.AI employees, according to a bankruptcy filing, and a judge still has to approve the deal.
• OpenAI reportedly published hundreds of mathematics papers written by a model it has not released.

Search takes a query of up to 128 characters and returns 1 to 50 results, 10 by default. Each result carries a snippet, a summary written for a model to read, a rerank score and an authority rating. Fetch takes one URL and returns the page title and a Markdown body. Both are capped at 5 queries per second per account, counted separately, and Fetch also has per-domain limits.

Only successful calls are billed, so failed, empty and timed-out requests cost nothing. DeepSeek is also making Search calls from Claude Code's Web Search tool and Codex's web_search free for the first two weeks. Agent and retrieval apps built on DeepSeek can now get live web content without a separate search provider. The endpoints arrived the same day as V4.1-Flash-1017, the model update covered on 17 October.

Kyutai, working with Inria and École des Ponts, released REKI on 17 October. It is a vision-language-action model, so it takes camera images and an instruction and outputs robot arm movements. It pairs a 2.2 billion parameter Qwen3.5-2B vision-language model with a 0.2 billion parameter action expert that generates motion using flow matching. Memory sits inside the action expert as 32 learned tokens per frame, cached for the last three frames. Kyutai reports that reading this cache adds 0.1 ms to a 21.8 ms action chunk on one H100, and that on RoboTwin it cut episodes where the arm oscillates from 18.6% to 4.1%.

In Kyutai's comparisons, REKI scores 90.4% on LIBERO-Plus, ahead of Qwen-RobotManip at 89.0% and π0.5 at 85.7%. It reaches 65.8% on RoboTwin Clean2Rand, behind Qwen-RobotManip at 68.5%, and 38.4% on RoboCasa365, where Xiaomi-Robotics-1 scores 57.4%. The weights are open under a restricted license.

Meta released Muse Spark Open on 15 October. Meta reports about 816 billion total parameters with about 42 billion active per token, a context window of up to 1 million tokens, and input in text, images, audio and video. Meta reports 61.8 on JobBench against 61.4 for GLM 5.3, and 76.4 on Terminal-Bench 2.1, where Kimi K3 scores 85.0. The model card lists an Artificial Analysis Intelligence Index of 39, below GLM 5.3 at 45. Meta says it worked with Apple so the model runs locally with MLX on two M5 Ultra Mac Studios, and the license is restricted.

• Disney agreed to hire more than 60 Character.AI workers in a bankruptcy deal, and Bloomberg puts the pay at about $180 million over three years plus $30 million in retention awards, though other reports give different totals.
• Anthropic has given an outside AI watchdog what Ars Technica describes as "employee-like" access to the company.
• OpenAI reportedly released hundreds of mathematics manuscripts produced by an unreleased model, 722 according to one report.
• Anthropic reportedly published quantum circuits designed with Claude Mythos 5 that it describes as up to 39 times more efficient.
• Modal reportedly raised money at 15.75 times revenue.
• ModelBest, with Tsinghua and OpenBMB, released PilotDeck 2.0, an open-source operating system for agents.